This project provides a solution to a pressing challenge faced by the protein arm at Cargill. They required detailed insights the future of the ever-evolving boxed-beef market. You see, sales contracts can be signed months before shipment, but the USDA only provides signed contract prices for the first 3 weeks into the future. Hence, for the medium term (4-8 weeks out) the business is essentially blind.

The business already had an implemented solution, but it was known by them to be intractable, requiring days to run, and even then only providing a fraction of the insights they required. I considered whether this problem was best solved as an optimization problem, but ultimately, decided to reimagine the project using a probabilistic approach.

This solution combines forecasting future prices based on expected market conditions, simulating hundreds of possible future perturbations, and reinforcement earning to generate a recommended sales "velocity" for the upcoming 8 weeks. These insights are then used to guide sales managers on how much of their existing stock they should target to move each week, where higher "velocities" indicate weeks where they should be more willing to sell.

This is a multi-step process. The first portion forecasts the expected prices assuming nothing critical occurs in the market. The second calculates the price elasticity based certain factors. The third takes all inputs and simulates thousands of possible scenarios, ultimately providing the sales plan belonging to the statistically most profitable outcome. In practice this works quite well and has been estimated by the protein arm to increase profits by around one million dollars.

Regression

AI Pricing and Promotion Decomposition

The meat-based protein arm of Cargill had a problem. When they made decisions to change product prices, or run product promotions, they were no longer seeing the same success they used to. In fact, they were starting to see market-share fall. What they needed was a data-driven product.

I developed a sophisticated solution to inform the business of the optimal grocery store pricing strategy as well as the expected impact for different types of promotions the stores could run. By using this the business can simulate the expected impact of any decision they make, or decisions made by key competitors.

To do this I created an advanced regression model, based on econometric principles, which forecasts the volume sold for that each week based on a myriad of factors. The "secret sauce" is that most of these factors represent an important lever or dial the business or their competitor's control.

The simulation aspect of this solution leverages the extracted elasticity for our own price changes, the elasticity for competitors price changes on the sale of our products, the boost in sales we receive from a particular type of promotion, and many others.

Deployment of this solution, and incorporating it into the toolset used by the sales team, drove a significant increase in profits and I am currently overseeing the scaling of this into additional stores. I designed this to be a platform, and not a set of standalone solutions. This conscious architectural design choice means that only relatively small alterations are needed for moving it into new locations.

Classification

Financial Product Forecasting Solution

Cargill purchases produce from farmers in a multitude of ways. For example, farmers can show up unannounced and sell their crop on-the-spot at whatever the current market-price is, they can negotiate a static price months ahead of time, or choose other complex options. During these negotiations farmers also provide an estimate for the amount to be delivered, but based on many factors this is also subject to change. In a situation like this, the more insight Cargill has at their disposal the better they can negotiate and plan targeted marketing strategies.

After discussing the details at length with business stakeholders and sales managers I devised a sophisticated solution which reliably provides the required forecasts. These include the ikelihood that a farmer will arrive within a particular period, the amount of crop they are likely to bring if they arrive, and the financial product the farmer is most likely to select during contract negotiations. My solution enabled a successful targeted marketing campaign, leading to a notable 12% increase in closure rate.

For the insights such as likelihood to arrive and amount of crop to bring I ultimately found that a statistical modeling approach, as opposed to ML, was most effective. For the question of the financial product they prefer I built a hierarchical classification model. This first forecasts the most likely family of product they prefer, and then once identified further refines the recommendation by forecasting the most likely product within this family.

Generative AI

LLM Assisted Code Migration

Translating legacy code into newer systems, with updated syntax or even a newer language, is a growing problem across most large companies. Cargill is no exception. Migrating extremely important systems written in archaic languages is difficult, time-consuming, and extremely expensive. This project seeks to minimize this burden by Leveraging state-of-the-art LLM modeling strategies.

One of the first things I investigated was whether we can translate the code directly into the new language and system, leaving the developer to briefly assess the new code via visual inspection, run pre-existing robust unit tests, and address any minor differences found. However, I quickly found that the problem wouldn't capitulate this easily. In many cases the integrations with the new system were different enough that the LLM was not able to sufficiently understand the differences and the developer was finding no significant benefits as they had to check through every line regardless.

With this in mind I changed tactics and although I did still provide this "possible" code translation, I added additional insight to help the developer identify the portions of the code most likely to need careful examination. To do this I supplied a customized pseudocode which summarizes the original code in a human-readable format, while explicitly indicating the sections where it believes that integration changes are most likely.

Using this even if the translation makes a mistake, the developer doesn't need to search through every line to find them. This is ongoing, but so far this hybrid approach has worked very well and we are on-track to shave almost one-million dollars from the estimated migration cost.